Some polyphony tests

My computer is a Dell Precision 3431 with Windows 10 Pro, i9-9900 8 core 16 threads 3.1 GHz base speed, 64 GB RAM.

*** THESE TESTS WERE INADVERTENTLY DONE WITH CPU LOCKED AT THE BASE SPEED OF 3.1 Ghz.
*** PLEASE SEE POSTS BELOW STARTING AT JAN 22, 2022 FOR TESTS WITH TurboBoost WORKING.

Its static polyphony is 10,000 voices at 48k sample rate, 1024 samples, higher definition pitch shifting on in HW 6. That's 20 notes held at 500 pipes per note on the 4+ GB polyphony test organ, hitting the top yellow bar on the CPU meter, just short of red. 7,500 voices (15 notes) stays in the green.

To test Nancy's polyphony demands, I used a 12" ruler covering 14 notes, some free fingers, and both feet to put as many long releases into play as possible using 18 note chords. This gave me about 50% more concurrent voices than I can do with both hands and feet but no ruler, so this is a bit of a stress test. Registration was all stops, all couplers. Amplifiers turned off! About 2 chords per second semi-legato to get long releases into play.

1. Nancy all 4 perspectives, all samples, no higher def pitch shifting, 48k sample rate, 16 bit samples to fit in 64 GB

Polyphony limit - polyphony meter bars - CPU meter bars (highest)
9000 - 2 yellow - 2 yellow
8000 - 2 yellow - 1 yellow
7000 - 2 yellow - in green
6000 - 2 yellow - in green

2. Nancy all 4 perspectives, all samples, with higher def pitch shifting, 48k sample rate, 16 bit samples to fit in 64 GB

Polyphony limit - polyphony meter bars - CPU meter bars (highest)
8000 - 1 yellow - 2 red
7000 - 2 yellow - 3 red
6000 - 2 yellow - 1 red
5000 - 3 yellow - 2 yellow
4000 - 3 yellow - in green

3. Nancy 2 perspectives (Front and Middle), all samples, with higher def pitch shifting, 48k sample rate, 24 bit samples

Polyphony limit - polyphony meter bars - CPU meter bars (highest)
8000 - 1 yellow - 1 red
7000 - 1 yellow - 3 yellow
6000 - 2 yellow - 2 yellow

Note: This is the configuration I play, with polyphony limit at 12512. Without the ruler abuse, the polyphony and CPU meters both stay in the green no matter how dense the music gets.

The required polyphony can be calculated approximately as follows:

Number of registers used * Reverberation time * Number of voices played simultaneously * Tone changes per second and voice

Example:
Number of stops used (40) * Reverberation time (6 s) * Number of voices played simultaneously (4) * Tone changes per second and voice (8)

= 40 x 6 x 4 x 8 = 7680 voices = required polyphony

Not knowing your formula to figure out the best polyphony with the Nancy sample, as it turns out after tweaking it for a few weeks, I am coming up with almost exactly the same setting you have…. I’m within 100 of your figure!

Thanks

Pat

Mark…

May I ask how many Stops you had engaged at the time of this test?

Pat

All 65 stops, all coupled and GO suboctave coupler. Amps off, so who cares if character reeds and celestes are in there?

However, the pedal stops were only played with 4 notes, while manual stops were played with 18 - 20 notes (14 note ruler, feet, spare fingers).

By the way, when trying to throw as many concurrent releases as possible, you have to hold each big chord long enough for long releases before restriking the keys. Fast playing uses less polyphony because you get short releases.

After another two months of playing with Nancy, a few updates:

1. I ended up preferring a mix of Close, Front, and Rear perspectives, about 40% Close, 60% Front, 50% Rear. To fit in 64 GB memory at 24 bits with all samples loaded, I truncated Close releases to 2 seconds at load time.

2. I added some Close because the Front ranks seem too buried in reverb for me. I want the long reverb, so don't want to truncate, and I really like the warmer sound of the Front mic position. But the reverb is so prominent that I have trouble hearing the attack and sustain through the release tails at times. Adding a fraction of Close ranks helps define the note starts and stops better. But I find close-recorded pipe sounds to be rather hard -- I feel it's like listening to a singer projecting to an audience, but I'm standing on their feet so they have to yell in my face. (ripping Spamelot)

3. I had been using 96k sample rate and 2048 samples before installing Nancy, but running Nancy at 3 perspectives at 96k was pushing the CPU at times, so dropped to 48k and 1024 samples. I've recently gone back to 96k and 2048, but had to turn off higher def pitch shifting on Nancy to run 3 perspectives.

4. So I needed a way to push the Front reverb slightly more into the background without loading the Close ranks. And with just Front and Rear loaded, using 96k with higher def pitch shifting won't overwork my CPU.

HW 6 makes this solution easy to try: Send a fraction of the Front rank signals to Perspective 2 and truncate releases on Perspective 2 in the Rank voicing screen. Route the new perspective to the same group as the original, full release Perspective 1. The attack and sustain part of the note gets amplified over the release part without losing any of the reverb length.

I wouldn't send more than 25 - 30% to the 2nd perspective because it's also amplifying the reverb buildup during the sustain sample, which would make a weird hump if too audible. ADDED: Actually works well at 50% so far.

I tested this method on Salisbury so far, where I also wanted a little more definition to hear note starts and stops without reducing reverb lengths. Works great so far.

Will update when I get to test it with Nancy. ADDED: This works and saves both memory and CPU while sounding better on my audio system.

Ein sehr interessanter Test, da ich selber diese Orgel kaufen möchte. Ich habe einen ähnlichen PC. Mir fehlen bei diesen Tests und verschiedenen Einstellungen die Latenzzeiten. Könnten sie diese noch angeben?

Lg Roglu

[mdyde]

Since this is an English-language forum, this is a Google translation into English of Roglu's post:

Ein sehr interessanter Test, da ich selber diese Orgel kaufen möchte. Ich habe einen ähnlichen PC. Mir fehlen bei diesen Tests und verschiedenen Einstellungen die Latenzzeiten. Könnten sie diese noch angeben?

A very interesting test, because I would like to buy this organ myself. I have a similar PC. I miss the latency times in these tests and various settings. Could you specify these?

[mdyde]

Thanks, Iain! (A copying/pasting error on my part.)

Since latency wasn't the purpose of my tests, I didn't record that information. A latency test would be designed a little differently. I was concerned with how much polyphony I could get out of a specific CPU without risk of audio problems due to CPU overload, with various audio quality settings and sampleset ranks.

Latency would vary with sample rate and buffer size, choice of audio interface and driver, and Windows performance tuning. But it doesn't vary much, if at all, with sampleset rank count, release duration, or the pitch shifting quality setting.

That is, you can check your latency with St Anne's and expect about the same on Nancy or Caen as long as the HW CPU meter stays in the green or low yellow range.

Sorry, I know you asked for the latency numbers, but I don't have time to do testing for that right now. Maybe when my new i9-12900K gets here I'll write up latency and polyphony tests for both machines.

Thank you for your quick response.

Just did a quick test to verify. HW's Sound delay ms (reported by the ASIO driver) is, unsurprisingly, constant between organs, whether idle or heavily loaded, with or without higher def pitch shifting. I checked using Nancy and St Anne's -- same results.

For my MOTU 24Ao, by sample rate and buffer size, the reported latencies are:

96k 512 11.7 ms
96k 1024 22.3 ms
96k 2048 43.7 ms

48k 512 22.3 ms
48k 1024 43.7 ms

No change when I push Nancy into the red on the CPU meter.

Well. I just noticed today that my i9-9900 was running at its base clock speed of 3.10 Ghz regardless of load, even though Turboboost was enabled in BIOS, while Speedstep was off to avoid lower clock speeds.

When I enabled Speedstep in BIOS just now and ran a polyphony test on Caen, the clock jumped up to 4.50 and I was able to set polyphony much higher. So I'll redo the tests and post corrected results here. (I also improved the testing method using MIDI recordings of test inputs.)

I don't understand this behavior. Speedstep should enable lowering the CPU speed below 3.10 Ghz when idle but should have nothing to do with raising it above 3.10, which Turboboost should enable. I wasn't expecting to have to enable both to get a higher clock speed. It seems I've had this CPU choked for two years.

I've just been looking up "Speedstep and Turboboost" - it seems you are not alone https://discuss.cakewalk.com/index.php?/topic/16946-intel-speedstep-and-my-dells-turning-it-on-resulted-in-better-performance/.

However, there posts which say that people have found that with speedstep enabled, Turboboost is disabled and performance is poorer.

This seem to depend on the motherboard, BIOS, and operating system.

Time to experiment (noting done working settings carefully before starting...)?

Iain